WSSIC-Net: Weakly-Supervised Semantic Instance Completion of 3D Point Cloud Scenes.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Semantic instance completion requires complete 3D object data and labels from partial 2.5D scans.
• Existing methods depend on full supervision, necessitating expensive and time-consuming ground-truth annotations.
• This reliance on extensive annotations limits practical applications of 3D scene understanding.

Purpose of the Study:

• To develop a Weakly-Supervised Semantic Instance Completion Network (WSSIC-Net) for 3D object completion from partial scans.
• To overcome the limitations of costly ground-truth data acquisition in real-world scenarios.
• To enable robust 3D shape recovery without complete object annotations.

Main Methods:

• WSSIC-Net utilizes 3D ground-truth bounding boxes, partial real-world objects, and unpaired synthetic 3D point clouds.
• A 3D detector encodes partial point clouds into features for supervised box prediction and weakly-supervised instance completion.
• A Generative Adversarial Network (GAN) completes partial object features using semantically consistent synthetic data.

Main Results:

• The proposed WSSIC-Net effectively performs weakly-supervised instance completion without needing complete 3D object ground truth.
• The integration of fully supervised 3D detection and weakly-supervised instance completion proved complementary.
• Evaluations on the ScanNet v2 dataset showed WSSIC-Net achieving performance comparable to state-of-the-art fully supervised methods.

Conclusions:

• Weakly-supervised learning offers a viable and cost-effective alternative for semantic instance completion.
• WSSIC-Net demonstrates the potential to significantly reduce annotation costs in 3D scene understanding tasks.
• The approach paves the way for more practical and scalable 3D object reconstruction in real-world applications.